Computers and Geotechnics 56 (2014) 202215

Contents lists available at ScienceDirect

Computers and Geotechnics

journal homepage: www.elsevier.com/locate/compgeo

Coupled hydro-mechanical fault reactivation analysis incorporating

evidence theory for uncertainty quantication
Leonardo C. Pereira a,b,c, Leonardo J.N. Guimares b, Bernardo Horowitz b, Marcelo Snchez a,
a
Zachry Department of Civil Engineering, Texas A&M University, USA
b
Department of Civil Engineering, Federal University of Pernambuco, Recife, Brazil
c
PETROBRAS Petrleo Brasileiro S.A., Brazil

a r t i c l e i n f o a b s t r a c t

Article history: The injection of water (or CO2) at high pressure is a common practice to enhance oil production. A crucial
Received 9 September 2013 component of this activity is the estimation of the maximum pressure at which the uids can be injected
Received in revised form 4 December 2013 without inducing the reactivation of pre-existing faults that may exist in the formation. The damage
Accepted 12 December 2013
zones typically formed around the geological faults are highly heterogeneous. The materials involved
Available online 3 January 2014
in the damage zones are characterized by the huge variation of their properties and high uncertainties
associated with them. To estimate the maximum allowable injection pressure this paper presents a novel
Keywords:
approach based on: a coupled hydro-mechanical formulation (for the numerical analyses); a criterion
Fault reactivation
Maximum injection pressure
based on the total plastic work (for the fault reactivation); and the evidence theory (for uncertainty quan-
Coupled hydro-mechanical analysis tication). A case study based on information gathered from an actual eld is presented to illustrate the
Evidence theory capabilities of the proposed framework.
Uncertainty quantication 2013 Elsevier Ltd. All rights reserved.

1. Introduction
Reservoir seals are natural traps that prevent the upward
migration of hydrocarbons. Fault reactivation may induce seal rup-
ture with the possibility of connecting the reservoir with other
rock blocks and nally with the surface. This process may lead to
oil release and the associated environmental accident. Fault reacti-
vation can be triggered by different factors, including the injection
of uids (i.e. H2O; CO2) at high pressure generally used to enhance
oil production. Injection of uids at high pressure represents a cen-
tral activity around oil production strategies. Perhaps one of the
biggest challenges of engineers and geologists involved in H2O/
CO2 injection projects is to estimate the maximum allowable injec-
tion pressure that will not hampered the seal condition of the fault.
In this work a methodology based on coupled numerical analysis
and uncertainty quantication (using evidence theory) is proposed
to estimate the maximum injection pressure in oil reservoir
systems.
The appropriate representation of a geological fault in a
numerical model requires a good understanding of the so-called
damage zones. Damage zones are formed by fractured rocks
associated with the deformation process near the less permeable
Corresponding author. Address: Zachry Department of Civil Engineering, Texas
A&M University, 3136 TAMU, College Station, TX 77843-3136, USA. Tel.: +1 979 862
6604; fax: +1 979 862 7696.
E-mail address: msanchez@civil.tamu.edu (M. Snchez).
fault core. A number of recent modeling efforts have been made
to better understand the mechanical and ow behavior in the dam-
age zones. For example Goodarzi et al. [1] performed a determinis-
tic modeling for CO2 storage in Nisku aquifer in Wabamun Lake
area in Canada. They concluded that the injection in the Nisku
aquifer may cause small surface heave only, with (very likely) no
environmental impact associated with surface deformations. The
hydro-mechanical behavior of a fault running across a reservoir
during a CO2 injection scenario was modeled by Ducellier et al.
[2]. Particular attention was paid to the inuence of some model
parameters on the uid ow response through the fracture. They
showed that the parameter with the major inuence on the ow
along the fault is the permeability of the lling material, and the
following ones (in a lower extent): the reservoir permeability,
the permeability of the joint elements and the initial opening of
the joint elements (these last two are strongly related). The main
aim of this work [2] was not to estimate the injection pressure that
reactivate the fault, but the effects of the owdeformation cou-
plings on the behavior of the fault zone. Perera et al. [3] developed
a 2D model using the Comsol Multiphysics simulator to obtain the
optimum injection pressure in coal seam without breaking its
sealant. They concluded that during CO2 injection, the cap rock is
vertically deformed by CO2 pressure. Moreover, the maximum ver-
tical deformation of the cap rock and the CO2 spread rate along the
interface increase with the increment of the CO2 injecting pressure.
The problem of uid ow coupled with the deformation of the nat-
ural formation has been discussed extensively in the literature. For

0266-352X/$ - see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.compgeo.2013.12.007
 L.C. Pereira et al. / Computers and Geotechnics 56 (2014) 202215 203

Nomenclature

r. divergence operator RHO Specic rock weight

r total stress tensor K0 earth pressure coefcient at rest
b body force vector kh horizontal permeability
/ porosity kv vertical permeability
ql liquid density E Youngs modulus
Pl liquid pressure t Poissons ratio
ql Darcys ow IDZ internal damage zone
u solid phase velocity EDZ external damage zone
qs solid density C2 external damage zone cohesion
ev volumetric strain NIRES reservoir Poissons ratio
r0 effective stress tensor ERES reservoir Youngs modulus
m auxiliary vector AA2 external damage zone friction angle
FY yield surface NI2 external damage zone Poissons ratio
h lode angle E2 external damage zone Youngs modulus
J 2nd stress invariant NIRIG denes K0 = NIRIG/(1  NIRIG)
p0 effective mean stress KRES reservoir permeability
c0 effective cohesion KDAMAGE2 external damage zone permeability
/0 effective friction angle KCORE core permeability
ep plastic strain tensor CCORE core cohesion
ry stress tensor at yielding NICORE core Poissons ratio
Wc plastic energy dissipated per unit volume ECORE core Youngs modulus
/i initial porosity AACORE core friction angle
ki intrinsic initial permeability tensor

example, Kojic and Cheatham [4] presented a treatment of the
plasticity theory in porous media with uid ow. Huang et al. [5]
used the random nite element method to investigate the
inuence of the standard deviation on various measures of the
equivalent coefcients of consolidation.
A key characteristic of the fault reactivation problem is the huge
range of variation of hydraulic and mechanical parameters in this
highly heterogeneous zone. Furthermore, the lack of reliable
experimental data associated with the materials in the damage
zone is quite common. Therefore, a formal framework is essential
for handling the uncertainties associated with this kind of analysis.
Uncertainty quantication is the process of determining the effect
of input uncertainties on response metrics of interest. The input
uncertainties can be characterized as either aleatory uncertainty,
which has irreducible variabilities inherent in nature; or epistemic
uncertainties, which are reducible uncertainties resulting from a
lack of knowledge. Since sufcient data is generally available for
aleatory uncertainties, probabilistic methods are commonly used
for computing response distribution statistics based on input prob-
ability density functions. Conversely, for epistemic uncertainties
data is generally scarce, making the use of probability theory ques-
tionable and resulting sometimes in non-accurate predictions.
When dealing with epistemic variables, non-probabilistic methods
generally based on the specication of intervals data are more
appropriate.
In the analyses presented in this work, the parameters uncer-
tainties are mainly related to the lack of (or limited) information
about them. So, the uncertainty is not aleatory but epistemic
and, therefore, a framework based on the evidence theory (a
non-probabilistic method) has been adopted in this work to deal
with the uncertainties associated with the main variables involved
in this problem.
Uncertainty quantication has been an important topic in al-
most all the engineering areas. Probability theory based on the
Bayesian school has been widely used in the past (e.g. [6,7]); but
more recently fuzzy methods and evidence theory are regarded
as ones of the legitimate extension of classical probability theory
(e.g. [8,9]). Dodagoudar and Venkatachalam [10] presented an
alternative approach, called fuzzy point estimate method, to esti-
mate the reliability of a nite earth slope. The approach incorpo-
rates the degree of membership value associated with each of
the uncertain input parameters used in the proposed model.
Schweiger and Peschl [11] developed a methodology based on evi-
dence theory and compared the results obtained with this tech-
nique against random eld nite element analysis by means of
assessing the probability of failure of a simple slope. They showed
that in some cases both methods give very similar results but in
other cases important differences are observed, in particular for
small spatial correlation lengths where the averaging procedure
adopted is less accurate. That study concentrated on the mechan-
ical problem only, without considering the hydro-mechanical
(HM) coupling in the analysis.
This paper focuses on the evaluation of the maximum allowable
injection pressure to assist oil production in pressurized reservoirs
involving a fault zone. Deterministic, probabilistic and non-proba-
bilistic analyses have been performed based on a typical fault
reactivation problem. This kind of analysis is generally solved ana-
lytically or numerically and the criterion adopted to decide
whether the fault will be reactivated or not is based on the position
of the current stress state with respect to a yield surface or enve-
lope (e.g. Mohr Coulomb envelope). In this work, a fully coupled
formulation has been used in the numerical analyses and for the
rst time the evidence theory has been used to quantify the uncer-
tainties typically associated with this complex problem. Another
novel contribution of this work is the adoption of the total plastic
work criterion in the fault zone to estimate the maximum injection
pressure. A quality of this criterion is that the assessment of the
fault reactivation is based on the behavior in a certain volume
(near the fault) and not in isolated points, as it happens in other
criteria adopted in the pass (e.g. the ones based on the yield surface
criterion). One disadvantage of the evidence theory is the high
CPU-time demand generally associated with the numerical analy-
sis. In this paper we propose the use of Articial Neural Networks
(ANN) to reduce the computational cost.
The paper is organized as follows: rst, concepts related to the
fault zone and the fault reactivation processes are introduced.
204 L.C. Pereira et al. / Computers and Geotechnics 56 (2014) 202215
Then, the main equations of the model and relevant HM couplings
are briey described. Afterwards, the case study is analyzed using
probabilistic and non-probabilistic methods. Finally the main con-
clusions of this research are presented.
2. Fault reactivation process
The loss of the reservoir seal is one of the most critical situa-
tions in the oil industry. The uncontrolled inux and leak-out by
pre-existing or induced discontinuities in the overlying rock mass
can lead to huge economic loss and environmental impact. The
trap is the stratigraphic or structural feature that ensures the jux-
taposition of reservoir and seals such that hydrocarbons remain
trapped in the subsurface, rather than escaping to the surface
due to their natural buoyancy. Fig. 1 presents (schematically) a
structural view of a trapped reservoir.
Geological discontinuities, such as faults, are inherent in most
petroleum formations. A fault plane (or fault zone) is a discontinu-
ity in the rock mass, quite common in most sedimentary basins. It
may be formed, amongst other, by tectonism or halokinesis pro-
cesses characterized by temporal distortion of evaporites. Fig. 2
shows a typical disposition of materials in a fault zone. The fault
plane is an intensely fractured region basically formed by a central
core and the damage zones. This set of materials is called fault
rock. The core comprises a ller material located between the fault
planes. The ller is composed of ne particles sizes (mainly formed
due to shear deformations), which may or may not be diageneti-
cally cemented. In geomechanical studies, it can be assumed that
the fault core is initially impermeable. The damage zone encom-
passes the mass of deformed rocks around the fault surface, which
results from the initiation, propagation and build-up of slip along
faults (e.g. [12,13]). The damage zone may extend up to several
meters beyond the fault core.
Fractures in fault zones may develop in a variety of angles with
respect to the direction of the maximum principal stresses gener-
ated during the frictional sliding of the adjacent faulted blocks
[14,15]. Due to the presence of the fractures, this region may have
different mechanical properties respect to the host rock. These
fractures are considered sealed at the beginning of the analyses,
and they act as a barrier for the migration of uids. In fact, the seal-
ant condition of the fault in both regions (i.e. central core and the
damage zone) is an essential requirement for the existence of the
oil reservoir.
However this sealant condition of the fault may change due to
geological processes or human actions. For example, typical opera-
tions associated with oil production, such as H2O or CO2 injections,
generally result in the re-distribution of in situ stresses, which may
lead to the reactivation of nearby faults and subsequent slips [16].
The mechanical integrity of the reservoir depends on a number of
Fig. 1. Main components in a petroleum reservoir in a sedimentary basin.
factors, amongst others: shear and tensile strengths of the forma-
tion rock and changes in uid pressures (which in turn induce
changes in effective stresses). If the injection of uids (i.e. H2O or
CO2) at high pressures is such that the associated changes in effec-
tive stresses induce yielding, signicant stress re-distribution and
plastic deformations are anticipated in the reservoir and fault zone.
Those changes generally lead to the reactivation of pre-existing
discontinuities (and/or the formation of new fractures) and the loss
of the fault sealing capability. Once the fault is reactivated, uids
may migrate between blocks and also from the reservoir to
shallower depths through the reopened discontinuity. The fault
reactivation can also pose a problem for well casing integrity that
eventually intersects these reactivated regions.
From above, it can be concluded that this problem is character-
ized by the strong coupling between the mechanical and the ow
problems [17]. For a particular case study, the injection pressure
that may trigger the fault reactivation will depend on the coupled
hydro-mechanical processes described above and on the materials
properties involved in the analysis. The formal approach used in
this work to analyze this type of problems is presented in the next
section.
3. Mathematical formulation and computer code
The fully coupled formulation proposed by Olivella et al. [18]
and the associated nite element program CODE_BRIGHT [19] have
been adopted for the numerical analysis. This approach is able to
deal with non-isothermal multiphase uid ow in deformable por-
ous media. For sake of simplicity, isothermal conditions have been
assumed in this work. Furthermore, only one uid has been consid-
ered in the analysis. This decision has been taken to reduce the
number of constitutive laws (and related model parameters) in-
volved in the uncertainty quantication analysis. This study can
be also considered as the necessary preliminary step before per-
forming a multiphase analysis.
The formulation is composed of balance equations (established
for the whole porous medium) and constitutive equations (related
to the intrinsic materials behavior). In this work three balance
equations have been adopted as follows: momentum balance,
mass balance of water and solid. The specic constitutive equa-
tions are described below and the adopted model parameters for
the analyses are presented in Section 4. More details about the
mathematical formulation can be found elsewhere (e.g. [18,19]).
3.1. Balance equations
Inertia effects are not considered in this study; therefore
momentum balance is reduced to the equilibrium equation:
r:r b 0 1
 L.C. Pereira et al. / Computers and Geotechnics 56 (2014) 202215 205

External Internal
Damage Zone Damage Zone Core
Reservoir (EDZ) (IDZ) (~5 meters)
(~25 meters) (~15 meters)

Fig. 2. Main fault zones. Damage zones and fault core.

where r is the symmetric tensor of total stresses and b is the body The mechanical model presented above has been combined
force vector. with a criterion to predict fault reactivation. A ductile fracture cri-
Considering a porous medium saturated with a single uid (e.g. terion to predict the onset of fracture opening based on the total
water), the mass balance of liquid is represented by: plastic work was rst suggested by Gillemot [20]. This criterion
assumes that the crack will start (and will then propagate) when
@/ql
r:ql ql /ql u
_ 0 2 the plastic energy dissipated per unit volume reaches a critical va-
@t
lue (Wc). The energy absorbed per unit volume during yielding can
where / is the porosity and ql is the liquid density. The liquid is be calculated as [20]:
considered compressible and its density depends on the liquid pres- Z
sure Pl (e.g. [19]). The Darcys ow is dened by ql and u is the solid Wc rY : dep dV 8
phase velocity.
The mass balance of solid can be expressed as: where ep is the tensor of plastic strains and ry is the stress tensor at
yielding. This study proposes to use the total plastic work criterion
@
1  /qs  r:1  /qs u
_ 0 3 to dene the maximum injection pressure. In this work the plastic
@t
work has been calculated for the volumes associated with each
where qs is the density of the solid phase, which depends on solid one of the materials involved in the damage zone (i.e. the plastic
grains compressibility. Applying the concept of material derivative, work integrated (Eq. (8)) for the volumes comprising: the core,
the variation in porosity can be expressed by [18]: the external and internal damage zones). The plastic work values
calculated in that way have been used to evaluate the maximum
D/ 1  / Dqs
1  /ev 4 injection pressure (as explained in Section 4). Such kind of criterion
Dt qs Dt
is recommendable because the prediction of the mobilization of the
where ev is the volumetric strain. geological fault is not based on what happens in isolated points, but
on the plastic behavior of the materials in the fault zone.
3.2. Constitutive equations As for the hydraulic problem, this work assumes that the intrin-
sic permeability depends on porosity through the following expo-
It is assumed that the mechanical behavior is controlled by the nential law [21] as follows:
effective stresses (r0 ), dened by: k ki eb//i 9
r_ 0 r_  mpl 5 where /i is the initial porosity, ki is the intrinsic initial permeability
where r_ is the total stress tensor and m is an auxiliary vector tensor and the parameter b controls the effect of porosity on
(mT = 1, 1, 1, 0, 0, 0). The mechanical constitutive model adopted permeability.
in this work is based on the DruckerPrager criterion, where the
yield surface (FY) can be expressed as: 3.3. Computer code
 
c0 0 The nite element program CODE_BRIGTH solves the equations
FY J  0 p Gh 0 6
tan / described above (i.e. mass balance, momentum balance, and con-
where G is a function of the lode angle (h): stitutive equations) in a fully coupled way. One unknown is asso-
p ! ciated with each one the equations presented above (i.e. u is
2 3 sin /0 associated with (1); liquid pressure is associated with (2); and
Gh 30 7 (4) is used to update /). It is a 3D code that adopts the Newton
3  sin /0
Raphson method to solve the non-linear problem and nite differ-
where J is the 2nd stress invariant of deviatoric stress tensor, p0 is ences to solve the evolution in time. The code has been extensively
the effective mean stress, c0 is the effective cohesion and /0 the validated in different simulations involving coupled geomechani-
effective friction angle. cal problems (e.g. [19,22,23]).
206 L.C. Pereira et al. / Computers and Geotechnics 56 (2014) 202215

4. Numerical modeling of a fault reactivation problem
The numerical modeling of problems associated with fault reac-
tivation has received increasing attention in the last few years (e.g.
[2426]). As mentioned in Section 2, the analyses associated with
the reactivation of geological faults in oil reservoir are generally re-
lated to the injection of uids (i.e. H2O or CO2) at high pressure.
Numerical modeling in this area is aimed at estimating the maxi-
mum injection pressure that can be applied to assist oil production
without inducing the reactivation of pre-existing faults or the gen-
eration of new fractures.
The numerical analyses presented in this paper focus on the
behavior near a fault zone. Fig. 3 shows a typical two-dimensional
geological cross-section associated with an oil reservoir. The rect-
angle in this gure delimits the domain studied in this paper.
The problem has been solved assuming 2D-plain-strain condi-
tions, with a rectangular domain 1560 m (wide) and 750 m (deep),
as shown in Fig. 4. In this problem, the fault is quite long and with
rather constant gradient, so 2D-plain-strain conditions are appro-
priate. The reservoir is a 50 m thick consolidated sandstone, over-
laid by a 400 m thick deposit of shale. The sea bed is at 130 m
Fig. 3. Typical 2D geological cross-section. The red square indicates the area of
inuence of the model analyzed in this study. (For interpretation of the references
to color in this gure legend, the reader is referred to the web version of this
article.)
Fig. 4. Geometry of the model: the reservoir is a 50 m thick consolidated sandstone
embedded 400 m and being located in a 130 m water depth region.
below the water level. The adopted boundary conditions are shown
in the same gure. The initial stress state was dened assuming
geostatic conditions, considering specic rock weight (RHO) equal
to 2300 kg/cm3 and K0 equal to 0.45.
The geological fault is represented by three materials: core,
internal damage zone and external damage zone (Fig. 2). The case
study is based on an actual eld located in Campos Basin, Rio de
Janeiro, Brazil. The properties adopted in this model were obtained
from both: log results (to dene ow properties); and triaxial tests
(to determine mechanical properties). The reservoir average hori-
zontal intrinsic permeability (kh) is equal to 50 milidarcy (i.e.
4.934  1014 m2) and the average porosity is equal to 0.2. Follow-
ing common practice in reservoir simulation, the vertical perme-
ability (kv) has been estimated according to the following ratio:
kh/10 = kv. Based on laboratory tests, the elastic properties (i.e.
Youngs modulus E and Poissons ratio m ) presented in
Table 1 have been adopted for the reservoir, overburden and
underburden rocks. The mechanical properties associated with
the fault zone (i.e. core, internal damage zone and external damage
zone) were estimated through 1st and 2nd after rupture (AR) tests
[27]. The main mechanical properties (i.e. E; m; c0 ; and /0 ) are also
presented in Table 1. It is worth mentioning that the uncertainty
associated with these parameters is high and the adopted values
for the deterministic case presented in this section correspond to
best estimate values.
The nite element mesh is composed of 6335 nodes and 12,550
triangular quadratic elements. Fig. 5b shows the adopted nite ele-
ment mesh. Fig. 5a presents a detail of the fault region where it can
be observed the high mesh renement adopted there, anticipating
the high gradients to be expected in this part of the domain. Fig. 5c
illustrates the adopted physical model associated with the damage
zones used to dene the numerical model. The fault zone has the
following characteristics: a core 5 m thick; an internal damage
zone 15 m thick (at each side of the core); and an external damage
zone 25 m thick. The core generally tends to be more rigid and less
permeable than the adjacent damage zones.
It has been assumed that an injection well is located at the left
of the fault zone and a production well is located at the right
(Fig. 4). This well disposition corresponds to a plausible scenario
in this kind of problem. It has been considered that injection well
operates with a well bottom hole pressure (BHP) up to 3.8 MPa
above the original pressure. It has been assumed that the produc-
tion well operates with a delta pressure of 4.0 MPa.
After some time pressurizing the reservoir, the effective
stresses will reach the yield criterion, jeopardizing the entire
mechanical stability of the formation and leading to the reactiva-
tion of the fault zone. The development of (dilatant) plastic
strains facilitates the migration of uids into the fault zone;
which would compromise the sealing capabilities of this zone.
If this happens, uids will migrate from the reservoir up to the
surface. The challenge is to estimate the maximum allowed
increase in pore pressure that will not trigger the exudation
process. The follow gures illustrate the process of fault
reactivation explained above.

Table 1
Mechanical and ow properties for each region considered in the model.

Horizontal permeability (mD) Porosity Young modulus (GPa) Poisson ratio Cohesion (MPa) Friction angle ()
Reservoir 50 0.2 30 0.3
Overburden 1e5 0.01 42 0.37
Underburden 1e5 0.01 26 0.26
Fault
Core 1e5 0.1 8 0.3 1.0 14
IDZ 1e5 0.2 8 0.3 0.8 16
EDZ 1e5 0.3 6 0.25 0.8 16
 L.C. Pereira et al. / Computers and Geotechnics 56 (2014) 202215 207

Fig. 5. Mesh adopted for the numerical analysis: (a) detail of the mesh at the fault zone; (b) mesh for the whole analyzed domain; and (c) illustration of the adopted physical
model associated with the damage zones used to dene the numerical model.

Fig. 6. Typical results associated with uid pressure distribution in the reservoir: (a) beginning of the reservoir pressurization; (b) just before the fault reactivation
phenomenon; and (c) just after the fault reactivation.

Fig. 6 shows typical results associated with changes in pore The signicant changes in the stress states may lead to yielding
pressure in the fault zone for the case that there is a lack of control and the development of plastic strains. Fig. 10 shows the concen-
on reservoir pressurization. Fig. 6a shows the beginning of the tration of shear strains in the fault zone.
reservoir pressurization. Fig. 6b presents the adopted scale for con- As mentioned before, a criterion based on the total plastic work
tours plots. Fig. 6c shows the pressure distribution in the reservoir (8) has been adopted in this work to dene the maximum injection
before the fault reactivation phenomenon. Fig. 6d illustrates the pressure. Fig. 11 shows the evolution of the plastic work in the
pressure distribution after the fault reactivation. three different materials of the fault zone (i.e. core, internal dam-
Fig. 7 shows the uid ow (module) in the area under study age zone and external damage zone). The maximum injection pres-
after fault reactivation. It can be observed that leakage takes place sure is dened when some of those materials reach a maximum
predominantly in the internal damage zone.
The reactivation of faults may induce important changes in the
displacement eld. The associated subsidence can damage seabed
infrastructure and (if applicable) surface facilities. Localized defor-
mations can induce the collapse of well casing located close to the
fault planes. Fig. 8 shows the displacement (module) eld pre-
dicted by the model after the fault reactivation process.
Once the activities associated with the exploitation of the reser-
voir start, the changes in the pore pressure (i.e. Fig. 6) modify the
initial stress state. Fig. 9 shows the concentration of shear stress
in the damage zone. As expected, a concentration of stresses is pre-
dicted near the fault plane; which is mainly induced by the con-
trast between the stiffness of the different materials involved in
the analysis. Fig. 7. Module of uid ow in the area under study just after fault reactivation.
208 L.C. Pereira et al. / Computers and Geotechnics 56 (2014) 202215
Fig. 8. Displacements eld just after fault reactivation: (a) horizontal displacement; (b) vertical displacement; and (c) displacement module.
Fig. 9. Distribution of shear stresses in the damage zone.
Fig. 10. Distribution of plastic shear strains in the damage zone.
value of the total plastic work equal to 102 [kg(m/s)2]. It can be
seen (Fig. 11) that, after this value is attained, the plastic work in-
creases signicantly, so it can be assumed that the fault is
reactivated.
For this specic case, the maximum allowed pore pressure
increment obtained was 3.2 MPa. However this result is related
to the particular parameters adopted in this analysis. A more gen-
eral study accounting for the uncertainties associated with the dif-
ferent parameters involved in this kind of problems is presented
later on the paper (Section 7). Next section presents some back-
ground information related to uncertainty quantication.
5. Uncertainty quantication
Uncertainty Quantication (UQ) is becoming an essential com-
ponent of analyses associated with complex problems dealing with
limited information. The uncertainty dual nature is described by
Helton [8] as:
 Aleatory uncertainty is the uncertainty that results from the fact
that a system can behave randomly. It is also known as:
stochastic uncertainty; Type A uncertainty; or irreducible
uncertainty.
 Epistemic uncertainty is the uncertainty that results from the
lack of knowledge about a system and/or its properties. It is also
known as: subjective uncertainty; Type B uncertainty; or reduc-
ible uncertainty. Evidence theory is one of the possible
approaches for dealing with epistemic uncertainty, as it is
explained later on.
Traditionally, the probability theory has been used to analyze
problems involving these two types of uncertainties. It is known
that the probability theory is a proper framework to quantify
uncertainties when dealing with aleatory variables. However, re-
cent contributions have shown that probabilistic methods may
not be the more convenient ones for analysis involving epistemic
uncertainties [8]. Other alternative methods, based for example
on evidence theory, seem more suitable for this kind of analysis [8].
Probabilistic analysis requires information about the probable
events. When this information is not available, the uniform
distribution function is generally used based on the Principle of
Insufcient Reason [28]. This principle can be interpreted as fol-
lows: in simple events, where the probability density function is
not known, it can be assume that the data is equiprobable. This
assumption is very appropriate for random variables, but when
dealing with epistemic ones, it is perhaps too approximate.
The evidence theory is an alternative to the probability theory
when attempting to quantify epistemic uncertainties (i.e., when
model input variables are considered reducible). The potential ben-
et of the evidence theory is that it allows a less restrictive descrip-
tion of the uncertainty, when compared with probability theory.
The uncertainty quantication through the evidence theory is com-
putationally more demanding, than the probabilistic method [8].
The additional computation effort comes mainly from the numer-
ical optimization process needed to compute the belief and plausi-
bility functions (details in Section 6.1). This paper proposes the use
of ANNs to speed up this process, as it is explained later on.
A further assumption in classical probability theory is inherent
in the additivity axiom. This axiom implies that information about
a particular event is associated with the knowledge of the comple-
ment of this event (i.e. the probability associated with the occur-
rence of a given event also informs about the probability of no
occurrence of this event [8]). Although the additivity assumption
and the principle of insufcient reason may be appropriate when
modeling random events associated with random uncertainty,
 L.C. Pereira et al. / Computers and Geotechnics 56 (2014) 202215 209

Fig. 11. Evolution of the plastic work in the three different materials of the fault zone.

these restrictions are questionable when dealing with situations input parameters without running the full factorial design. The
involving epistemic uncertainty [8]. Examples of these types of sit- subset is chosen so as to exploit the sparsity-of-effects principle
uations are: (a) little information about the model input parame- to expose information about the most important features of the
ters is available; (b) the available information is ambiguous or fault reactivation problem, while using a fraction of the effort
contradictory. required in a full factorial design in terms of experimental runs
In cases where the available information is not enough to quan- and resources. The Table 2 presents some of the 15 variables
tify the uncertainty with a probability function, it is sensible to considered in the sensitivity analysis. The Monte Carlo numerical
consider a possible variation range of the variable dened by inter- simulations show that eight parameters were the most inuential
vals, or sets of possible values. The denition of a range (or a set) of ones in terms of the maximum injection pressure. The tornado plot
probable values associated with a given variable implies that the presented in Fig. 12 shows the impact of the main variables on this
Principle of Insufcient Reason is not enforced. A priori, probabili- problem according to the sensitivity analysis.
ties can be established to sets, without having pre-established The vertical axis presents the variables considered in this study
assumptions about the probabilities of the individual events. Fur- and the horizontal axis represents the weighted inuence of those
thermore, the additive axiom is not enforced, therefore the com- variables on the nal result. The sum of those coefcient for all the
plementary sets of probabilities measures do not need to sum 1 variables considered in the analysis is equal to one. The sign is dic-
(when this occurs, this framework converges to the traditional tated by the way in which this variable impact on the nal result.
probabilistic representation) [8]. For example, the relative weight of C2 (cohesion of the external
According to Helton et al. [29,30], there are three main theories damage zone) is +0.3. This means that an increase of C2 is associ-
from which the uncertainty representation has been approached ated with an increment (due to the + sign) of the maximum allow-
by intervals: (i) imprecise probabilities; (ii) possibility theory; able injection pressure and the relative impact (when compared
and (iii) evidence theory, also known as DempsterShafer theory with the other variables) is 30%. A negative sign implies that an in-
(e.g. [3133]). This study adopts the evidence theory to quantify crease in the value of this variable is related with a decrease of the
the uncertainties associated with the fault reactivation problem. maximum injection pressure. The other variables presented in this
The main motivations behind the selection of the evidence theory plot are as follows NIRES (the reservoir Poissons ratio); ERES (the
are: (a) it is a sound theory; (b) it is easy to correlate evidence the- reservoir Youngs modulus); AA2 (the friction angle of the external
ory and the probability theory; (c) the large number of successful
applications reported in the last few years (e.g. [3436]); and (d)
the versatility of the theory to represent and combine different Table 2
types of evidences obtained from multiple sources. More informa- Mechanical and ow properties considered in the sensitivity analysis.
tion about evidence theory can be found elsewhere [3739]. Property Symbol Uncertainty
Characterization
Reservoir Poissons ratio NIRES Uniform (0.20.4)
6. Uncertainty quantication in fault reactivation Reservoir Youngs modulus (GPa) ERES Uniform (2050)
External damage zone /0 () AA2 Uniform (1430)
External damage zone cohesion C2 Uniform (0.81.2)
In this work, typical results obtained from a UQ study using the
(MPa)
well-known probability theory are compared against the results External damage zone Poissons ratio NI2 Uniform (0.20.4)
obtained from the analyses based on the evidence theory. The case External damage zone Youngs E2 Uniform (510)
presented in Section 4 is used in the analyses. In complex problems modulus (GPa)
(like this one) involving a large number of variables, it is highly Denes K0 = NIRIG/(1  NIRIG) NIRIG Uniform (0.30.34)
Specic rock weight (kg/m3) RHO Uniform (22002600)
recommended to perform a sensitivity analysis to identify the
Reservoir permeability (mD) KRES Uniform (2080)
key variables that will control the problem [4042]. The sensitivity External damage zone k (mD) KDAMAGE2 Uniform (1e52e5)
study will allow focusing the uncertainty analysis on those vari- Core permeability (mD) KCORE Uniform (1e52e5)
ables that have a major impact on the model response. In this work Core cohesion (MPa) CCORE Uniform (0.20.8)
Core Poissons ratio NICORE Uniform (0.20.4)
the sensitivity analysis has been based on the Monte-Carlo exper-
Core Youngs modulus (GPa) ECORE Uniform (510)
imental design technique proposed by Martens et al. [43]. Using Core /0 () AACORE Uniform (1430)
this kind of approach it is possible to reveal the more inuential
210 L.C. Pereira et al. / Computers and Geotechnics 56 (2014) 202215

Fig. 12. Tornado diagram showing the results of the sensitivity analysis of the most inuential parameters respect to the pore pressure increment.

damage zone); NI2 (the external damage zone Poissons ratio); E2
(the external damage zone Youngs modulus); RHO (the specic
rock weight, used to dene the initial vertical stress); KRES, KDAM-
AGE2 and KCORE are the permeabilities of the reservoir, damage
zone 2 and core zones, respectively. NIRIG denes the K0 (i.e. the
coefcient of earth pressure at rest is used to dene the initial hor-
izontal stress). Fig. 12 presents the eight more inuential ones only
(i.e. with a relative impact of |0.1%| or higher), plus the three per-
meabilities mentioned above.
Interestingly, the core and internal damage zone properties
were not very inuential on the model response. This is because
the external damage zone (initially sealant) controls the failure
of the fault zone. It is also remarkable that some variables with
high uncertainties (such as permeability) have a minor impact on
the maximum pore water pressure that can be injected. It can be
observed (Fig. 12) that the impact of the permeability of the differ-
ent materials is around 0.02%. This can be attributed to the fact that
the maximum injection pressure that may induce the fault reacti-
vation is mainly controlled by the mechanical problem. Variations
on the permeability values may have an impact on the time at
which the fault reactivation process may start, but its inuence
on the maximum predicted injection pressure is rather marginal.
The sensitivity analysis presented above was used to select the
variables to be included in the uncertainty quantication study
using both probabilistic and evidence theories.
There are different alternatives to combine evidence [29]. Typ-
ically, they can be subdivided in 4 groups: (i) consonant evidence,
when multiples sources agree in all subsets ranges; (ii) consistence
evidence, where at least there is one element that is common to all
subsets; (iii) arbitrary evidence, there is no element common to all
subsets, but some subsets may have elements in common; and (iv)
disjoint evidence, where any two subsets have no elements in com-
mon. In this wok, it was assumed arbitrary evidence based on ex-
perts knowledge. Geologists and reservoir engineers working in
this topic were asked to dene the ranges of each uncertainty
parameter and also the weights of each interval. Using the Demp-
sters rule of combination [32], the weights (or basic probability
assignment, BPA) presented in Fig. 13 were obtained. This informa-
tion is used in the evidence theory to calculate the belief and plau-
sibility functions as explained in the following sections.
Uniform density functions were assumed for the analysis using
the probability theory. Fig. 13 shows that the same values of the in-
put parameters are adopted when applying the probability and evi-
dence theories. However, it is different to say that a given model

Fig. 13. Range of variation of the different model parameters incorporated in the uncertainty quantication analysis. For the case of evidence theory the weight are also
presented. This data was dened after experts opinions.
 L.C. Pereira et al. / Computers and Geotechnics 56 (2014) 202215 211

X
parameter falls in the interval [a,b], than to say that the value has ~ 6 y
PlY y mEY A 13
an uniform distribution in [a,b]. A\Ay 0

At an intuitive level, belief is a measure of the amount of infor-

6.1. Belief and plausibility functions
mation that indicates that a statement is true, and plausibility is a
measure of the amount of information that indicates that a state-
The rst work in evidence theory was done by Shafer [32] as an
ment could be true.
expansion of Dempster [31] approach. In a nite discrete space, the
If one considers the universe of all probability distributions that
evidence theory or DempsterShafer theory (DST) can be inter-
are consistent with the focal elements and BPAs that dene an evi-
preted as the generalization of the probability theory for the case
dence space on a set XE. Then, for a subset U (i.e. y ~ 6 y) of X, Bel(U)
in which the probabilities are assigned to sets of values. In
is the smallest possible probability that could occur for U over the
traditional probability theory, evidence is associated with only
set of indicated distributions and Pl(U) is the largest possible prob-
one possible event. In the DST, the evidence may be associated
ability that could occur for U over the set of indicated distributions.
with multiple events (i.e. sets of events). As a result, the DST evi-
Central to this is the idea that all probability distributions that do
dence is presented with a higher level of abstraction, i.e. without
not violate the assumed properties of the evidence space are under
inferring a prior probability distribution of the input variables.
consideration.
When evidence is sufcient to assign probabilities to single events,
These concepts can be easily explained using the simple exam-
the DempsterShafer model falls in traditional probabilistic
ple as follows. Suppose that for the proposition that says the fault
formulation.
is reactivated there are a belief of 0.5 and a plausibility of 0.8. This
One of the most important features of the evidence theory is
means that the existing evidence allows to strongly state that the
that the model is designed to handle different levels of accuracy
proposition is true with a condence of 0.5. However, the evidence
(i.e. regarding the available data), and no further assumption is
contrary to that hypothesis (i.e. the fault is not reactivated) only
needed to represent them. It also allows the direct representation
has a condence of 0.2. The remaining mass of 0.3 (i.e. the gap be-
of the uncertainty responses of the system. For example if the input
tween the 0.5 supporting evidence on the one hand, and the 0.2
variable is imprecise, it can be characterized by a set or range of
contrary evidence on the other) means that the fault could either
values and, consequently, the resulting output will also be a set
be reactivated or not. This interval represents the level of uncer-
(or a range of values).
tainty based on the evidence in the system.
Two core concepts in evidence theory of Basic Probability
The denition of the belief and plausibility functions for the
Assignment (BPA) and focal elements are introduced as follows:
model output has been done through an optimization process.
(i) the BPA for a set is the amount of probability that can be as-
The main component of the optimization analysis is presented in
signed to that set but cannot be decomposed into additional prob-
the following section.
abilities for subsets of that set; and (ii) focal elements are those
sets that have nonzero BPAs. This then leads naturally to a clear
distinction between an evidence space for a set XE of possible out- 6.2. Optimization algorithm to compute CPF and CBF
comes and a probability space for a set XP of possible outcomes.
Suppose a model represented by: The belief and plausibility functions are dened for the input
variables, x. For each input variable it is necessary to specify inter-
y f x 10
vals (and the associated bounds) and Basic Probability Assign-
where, x = [x1, x2, x3, . . . , xnX] is the vector of the input variables and ments (BPA). A brief description of the necessary steps to
y = [y1, y2, y3, . . . , ynY] is the vector associated with the model re- compute these two functions of the response is given below (based
sults. The probability theory requires that uncertainty associated e.g. in the data presented in Fig. 13).
with x should be dened by a density function (e.g. normal, triangu-
lar or uniform). To characterize the uncertainty in x is necessary to (1) Combine intervals of each input variable into combination
establish a probability space X P ; X P ; mPX where XP is the sample cells (for example, if a case has 2 input variables and each
space of possible values of x; X P is the set of subsets of XP chosen of these is dened by 3 intervals, the number of combination
appropriately (r-algebra); and mPX is a probability factor [29]. cells is equal to 9). The total number of such cells is given by:
Thus, the uncertainty in y is estimated by a cumulative distribu-
tion function (CDF). niv
Y
Z X
nS ncell ninti 14
~ 6 y
proby df xjydX dX df xjy=nS 11 i1
XP i1
where: ncell = total number of combination cells, nint(i) = number of
where y ~ is the expected value; nS is the number of samples; intervals of input variable i, niv = number of input variables. The ba-
d(f(x)|y = 1) if f x 6 y and d(f(x)|y) = 0 if f(x) > y. sic probability assigned to each cell is the product of the BPAs of the
Conversely the evidence theory does not impose a rigid struc- involved intervals.
ture on the uncertainty characterization. The evidence theory is a (2) Perform the realizations involving all the possible combina-
more exible approach. To characterize the uncertainty in x it is tions dened above. For each cell solve the following two
necessary to dene an evidence space: X P ; X P ; mPX X E ; X E ; mEX optimization problems have to be solved to nd the lower
where XP = XE is the sample space of possible values of x but (lb) and upper (ub) bounds of the response:
X P X E . In probability theory, if a subset A belongs to X P the prob-
lb DPminj Minimize DPx 15
abilities associated to its complement Ac can be obtained from the
additive axiom (i.e. the complementary sets of probabilities have
to sum 1); while in evidence theory, the uncertainty in y is esti- ub DPmaxj Maximize DPx 16
mated by two uncertainty functions: belief and plausibility func-
tions [29]. where DP(x) = reactivation pore pressure increment modeled
X through a previously trained neural network; (xl)j and (xu)j are the
~ 6 y
BelY y mEY A 12 bounds of the input variables intervals associated with combination
AAy
cell j. The adopted neural network is explained in Section 6.3.
212 L.C. Pereira et al. / Computers and Geotechnics 56 (2014) 202215
The ub and lb pressures obtained after the optimization process
(i.e. Eqs. (15) and (16)) are used to plot the plausibility and belief
functions. These results are not organized in a particular order.
They need to be organized in ascending order, as it is explained
below.
(3) Sort in ascending order the lower bounds of the output
response intervals obtained in Step 2. Once the data is orga-
nized, to obtain the plausibility function the lb pressure
(Step 2) calculated per each combination (Step 1) is plotted
against the corresponding cumulative probability. The belief
function is plotted in a similar way, except that one should
use the up pressures (also obtained in Step 2).
6.3. Neural network
The optimization process explained in Section 6.2 involves a
large number of realizations. Different options are available to
perform these simulations. A numerical code can be used
(e.g. CODE_BRIGHT for the analysis presented in this paper), or
alternatively, any properly calibrated proxy (capable of performing
the numerical analyses) can be used to reduce the high CPU-
demand typically associated with the optimization process
involving multiple variables. In this work an articial neuronal
network has been adopted as proxy.
Neural networks are composed of simple elements operating in
parallel. Commonly, neural networks are adjusted in the way that a
particular input leads to a specic target output. The network is ad-
justed based on a comparison between the output and the target
until these values become similar. Different types of ANNs have
been proposed for forecasting and modeling engineering problems
(e.g. [4447]). A feedforward neural network with back propaga-
tion was trained in this study. The ANN has one hidden layer of
Fig. 14. Main results related to the Articial Neural Network tting: (a) results of the training; (b) validation of the tting; (c) test results; and (d) the global tting.
10 sigmoid neurons followed by an output layer of linear neurons.
The linear output layer is most often used for function tting
(or non-linear regression) problems.
When training multilayer networks, the general practice is to
rst divide the data into three subsets. The rst subset is the train-
ing set, which is used for computing the gradient and updating the
network weights and biases. The second subset is the validation
set. The error on the validation set is monitored during the training
process. When the network begins to over t the data, the error on
the validation set typically begins to rise. The network weights and
biases are saved at the minimum of the validation set error. The
test set error is not used during training, but it is used to compare
different models. It is also useful to plot the test set error during
the training process. If the error on the test set reaches a minimum
at a signicantly different iteration number than the validation set
error, this might indicate a poor division of the data set.
As discussed earlier, 2000 coupled simulations had been per-
formed to create an extensive database. The ratios for training,
testing and validation are 0.7, 0.15 and 0.15, respectively. Fig. 14
presents the t for these three subsets and also the global tting:
(a) results of the training; (b) validation tting; (c) test results
and (d) the global tting. Based on these results, the performance
of ANN for this problem can be considered very satisfactory and,
therefore, this ANN can be used as a proxy to save CPU-time.
7. Results
This section focuses on the use of the evidence theory for esti-
mating the maximum injection pressure in problems related to
oil production. Considering the limited (and in some cases lack
of) information regarding many parameters involved in this kind
of analysis, it is highly recommendable to dene via an uncertainty
 L.C. Pereira et al. / Computers and Geotechnics 56 (2014) 202215 213

analysis the maximum pressure at which a uid can be injected in
a reservoir.
Fig. 15a presents the main results comparing the use of proba-
bility and evidence theories for uncertainty quantication associ-
ated with the fault reactivation problem presented in Section 5.
The cumulative distribution function (CDF, Eq. (13)) was obtained
after the application of the probability theory, computed according
to [6]. The evidence theory was used to compute the cumulative
belief function (CBF, Eq. (14)) and the cumulative plausibility func-
tion (CPF, Eq. (15)). Complementary cumulative distribution func-
tion (CCDF), complementary cumulative belief function (CCBF) and
complementary cumulative plausibility function (CCPF) were di-
rectly obtained from the CDF, CBF and CPF, respectively. Fig. 15a
shows the results in terms of probability of fault reactivation (i.e.
cumulative probability to be lower) and it should be used when
one wants to look at the probability or belief or plausibility that
a give increment of the injection pressure will reactive the fault.
Fig. 15b presents the outputs of the same analyses but in terms
of probability of fault reactivation (i.e. complementary cumulative
results, probability to be higher), and it should be used to analyze
the complementary case, that is, when one wants to check that the
fault will not reactivate for a given increment of the injection pres-
sure. The complementary cumulative results are more appropriate
when the likelihood of exceeding a value is under consideration.
In this work the activation criterion is a function of the total
plastic work (Eq. (8)). When in any of fault regions the plastic work
exceeds a given maximum value (set as a tolerance in this
analysis), it is considered that the sealing capability of the fault
is breached and any additional pore pressure increment could
trigger the fault reactivation process. For example, if the maximum
injection pressure obtained from the deterministic case (i.e.
Pl = 3.2 MPa, Section 4) is analyzed in the context of the results pre-
sented in Fig. 15, according to the classical probability theory, it
can be concluded that the probability of a fault reactivation when
the uid is injected at 3.2 MPa is around 0.20. However, according
to the evidence theory the probability of a fault reactivation can be
between 0.65 (as a maximum value) and 0.00 (as a lower bound).
According to the classical probability theory, the probability of the
fault to remain stable is 0.8 if an increment of the injection pres-
sure equal to 3.2 is applied, and the associated degree of belief ob-
tained with the epistemic theory is 0.35 (Fig. 15b). It can be seen
that the evidence theory provides more complete information to
assist the decision maker.
One can also use these plots to analyze the maximum pressure
that can be applied based on a given probability, for example lets
assume that 0.10 is the largest possible probability that an injec-
tion project manager can take respect to the reactivation of a fault.
It is important to remember (see Section 6.1) that, Bel(U) is the
smallest possible probability that could occur for U over the set
of indicated distributions and Pl(U) is the largest possible probabil-
ity that could occur for U over the set of indicated distributions.
Therefore, considering rst the CDF, the increment in the injection
pressure can be as higher as 3.0 MPa (i.e. intersection of 0.10 prob-
ability with CDF). Looking at the evidence theory (i.e. CPF and CBF

Fig. 15. Main curves obtained from the probability and evidence theories. (a) Cumulative distribution function (CDF); cumulative belief function (CBF) and cumulative
plausibility function (CPF). (b) Complementary cumulative distribution function (CCDF); complementary cumulative belief function (CCBF) and complementary cumulative
plausibility function (CCPF). CDF and CCDF were obtained from the application of probability theory. CBF, CPF, CCBF and CCPF were obtained from the application of the
evidence theory. Note that 3.2 MPa corresponds to increment of the injection pressure obtained from the deterministic analysis.
214 L.C. Pereira et al. / Computers and Geotechnics 56 (2014) 202215
curves), one may conclude that for the same given probability (i.e.
0.10) for pressure increments equal or lower than 2.0 MPa the fault
will reactivate with a plausibility of 0.10. Note that for that incre-
ment of injection pressure (i.e. 2.0 MPa) the degree of belief that
the fault will reactivate is equal to 0.0. Similar analyses can be done
looking at the complementary cumulative probabilities (i.e.
Fig. 15b). It can be interpreted that the design of the injection pres-
sure based on the combination of CBF and CPF curves will be, ini-
tially, conservative (and perhaps safer) when compared respect to
the ones based on CDF curve.
It is worth to mention that this has been a rst attempt in the
application of this methodology to assess the maximum injection
pressure in reservoirs. Obviously improvements can be introduced
in this kind of analysis, amongst others: establishing correlations
between the parameters adopted in the study; include the inu-
ence of more complex initial stress conditions; and performing
multiphase analysis involving two uids. In any case, the results
presented in this section are very promising and encourage the
use of this kind of approach to design the maximum injection pres-
sure when assisting oil production or storing CO2.
8. Conclusions
This paper focuses on the estimation of the maximum allowable
injection pressures such the mechanical integrity of a pressurized
reservoir is not compromised. In particular, the problem of fault
reactivation during oil production assisted by injection of a uid
at high pressure has been studied in detail. Relievable analysis in
this area should account for the inherent uncertainties associated
with the materials involved in these kinds of problem. In this work,
a framework based on the evidence theory has adapted to the par-
ticular condition of this problem. When dealing with epistemic
variables (as in this case) it has been shown that this methodology
is an excellent alternative to more traditional methods based on
probabilistic method. Probabilistic methods require the denition
of prior density functions related to the uncertainties associated
with the parameters contemplated in the numerical analysis. This
requirement may be difcult to meet in cases dealing with no (or
limited) information about these density functions. As shown,
the evidence theory only requires instead the denition set of val-
ues for the variables and the corresponding weighing factor.
The case analyzed in this paper is based on information gath-
ered from an actual oil production eld. A fully coupled hydro-
mechanical formulation together with a trained ANN have been
adopted for performing the analyses involved in the different
stages of a UQ study. Two distinct curves are obtained from this
analysis: the cumulative plausibility function (CPF) and the cumu-
lative belief function (CBF). In addition to the analysis based in the
evidence theory, a conventional probabilistic analysis and a deter-
ministic study (using the best estimate values for all the variables)
have been performed.
It can be concluded that information provided by the evidence
theory is more general and complete when compared against the
probabilistic and deterministic results. The proposed approach it
has been shown very reliable and promising for its application in
this type of problems. More complex analyses involving multi-
phase ow, correlation between parameters, 3D, and more com-
plex geological conditions are underway.
Acknowledgements
The authors acknowledge the nancial support from
PETROBRAS, CNPq (Brazilian National Research Consul) and
Foundation CMG (Chair on Reservoir Simulation at UFPE). It is also
gratefully acknowledged by the authors the assistance provided by
the Reservoir Geomechanics Group at PETROBRAS to perform this
work. The expert opinion of the members of this group has been
instrumental to develop the data base use in this study; special
thanks to (amongst others): Alvaro Maia da Costa, Luis Carlos de
Sousa Jr., Claudio Amaral, Claudio Lima and Erick Slis. Finally, we
would like to thank the reviewers of this paper for their comments
and suggestions.
References
[1] Goodarzi S, Settari A, Keith D. Geomechanical modeling for CO2 storage in
Nisku aquifer in Wabamun Lake area in Canada. Int J Greenh Gas Control
2012;10:11322.
[2] Ducellier A, Seyedi D, Foerster E. A coupled hydromechanical fault model for
the study of the integrity and safety of geological storage of CO2. Energy Proc
2011;4:513845.
[3] Perera MSA, Ranjith PG, Choi SK, Bouazza A. A parametric study of coal mass
and cap rock behaviour and carbon dioxide ow during and after carbon
dioxide injection. Fuel 2013;106:12938.
[4] Kojic MC, Cheatham JB. Theory of plasticity of porous media with uid ow. R
Soc Petrol Eng J 1974;14(3):26370 [Int J Rock Mech Min Sci Geomech Abstr].
[5] Huang J, Grifths D, Fenton G. Probabilistic analysis of coupled soil
consolidation. J Geotech Geoenviron Eng 2010;136(3):41730.
[6] Pugachev VS. Probability theory and mathematical statistics for
engineers. Pergamon Press; 1984.
[7] Blanchard BS, Fabrycky WJ. Systems engineering and analysis. Pearson
Prentice Hall; 2006.
[8] Helton JC, Oberkampf WL, Johnson JD. Competing failure risk analysis using
evidence theory. Risk Anal 2005;25(4):97395.
[9] Bai YC, Han X, Jiang C, Liu J. Comparative study of metamodeling techniques for
reliability analysis using evidence theory. Adv Eng Softw 2012;53:6171.
[10] Dodagoudar GR, Venkatachalam G. Reliability analysis of slopes using fuzzy
sets theory. Comput Geotech 2000;27(2):10115.
[11] Schweiger HF, Peschl GM. Reliability analysis in geotechnics with the random
set nite element method. Comput Geotech 2005;32(6):42235.
[12] Cowie PA. Growth of faults by accumulation of seismic slip. J Geophys Res
1992;97(B7):1108595.
[13] McGrath A. Damage zone geometry around fault tips. J Struct Geol
1995;17(7):101124.
[14] Petit JP. Criteria for the sense of movement on fault surfaces in brittle rocks. J
Struct Geol 1987;9(56):597608.
[15] Storti F, Salvini F. The evolution of a model trap in the central Apennines, Italy:
fracture patterns, fault reactivation and development of cataclastic rocks in
carbonates at the Narni anticline. J Petrol Geol 2001;24(2):17190.
[16] Chapman RE. Petroleum geology. Amsterdam, New York: Elsevier; 1983.
[17] Cappa F, Rutqvist J. Modeling of coupled deformation and permeability
evolution during fault reactivation induced by deep underground injection of
CO2. Int J Greenh Gas Control 2011;5(2):33646.
[18] Olivella S, Carrera J, Gens A, Alonso E. Nonisothermal multiphase ow of brine
and gas through saline media. Trans Porous Media 1994;15(3):27193.
[19] Olivella S. Numerical formulation for a simulator (CODE_BRIGHT) for the
coupled analysis of saline media. Eng Comput 1996;13(7):87112.
[20] Gillemot L. Criterion of crack initiation and spreading. Eng Fract Mech
1976;8(1):23953.
[21] Snchez M, Gens A, Guimares L. Thermalhydraulicmechanical (THM)
behaviour of a large-scale in situ heating experiment during cooling and
dismantling. Can Geotech J 2012;49(10):116995.
[22] Snchez M, Gens A, Guimares L, Olivella S. Implementation algorithm of a
generalised plasticity model for swelling clays. Comput Geotech
2008;35(6):86071.
[23] Gens A. A full-scale in situ heating test for high-level nuclear waste disposal:
observations, analysis and interpretation. Geotechnique 2009;59(4):37799.
[24] Rutqvist J. Status of the TOUGH-FLAC simulator and recent applications related
to coupled uid ow and crustal deformations. Comput Geosci
2011;37(6):73950.
[25] Neves MC, Paiva LT, Luis J. Software for slip-tendency analysis in 3D: a plug-in
for Coulomb. Comput Geosci 2009;35(12):234552.
[26] Gambolati G, Teatini P, Tomasi L. Stressstrain analysis in productive gas/oil
reservoirs. Int J Numer Anal Methods Geomech 1999;23(13):1495519.
[27] Costa AM, et al. Setting the maximum water injection pressures. PETROBRAS;
2007.
[28] Savage L. The foundations of statistics. J Consult Psychol 1955;19(3):23745.
[29] Helton JC, Johnson JD, Oberkampf WL, Storlie CB. A sampling-based
computational strategy for the representation of epistemic uncertainty in
model predictions with evidence theory. Comput Methods Appl Mech Eng
2007;196(3740):398098.
[30] Helton JC, Johnson JD, Oberkampf WL, Storlie CB. A sampling-based
computational strategy for the representation of epistemic uncertainty in
model predictions with evidence theory. SAND2006-5557. Albuquerque,
NM: Sandia National Laboratories; 2006.
[31] Dempster AP. Upper and lower probabilities induced by a multivalued
mapping. Ann Math Stat 1967;38(2):32539.
[32] Shafer G. A mathematical theory of evidence. Princeton University Press; 1976.
 L.C. Pereira et al. / Computers and Geotechnics 56 (2014) 202215
[33] Yager RR. Arithmetic and other operations on DempsterShafer structures. Int
J ManMachine Stud 1986;25(4):35766.
[34] Curcur G, Galante GM, La Fata CM. Epistemic uncertainty in fault tree analysis
approached by the evidence theory. J Loss Prevent Proc Ind
2012;25(4):66776.
[35] Khalaj M, Makui A, Tavakkoli-Moghaddam R. Risk-based reliability assessment
under epistemic uncertainty. J Loss Prevent Proc Ind 2012;25(3):57181.
[36] Wang P, Lu Z, Tang Z. A derivative based sensitivity measure of failure
probability in the presence of epistemic and aleatory uncertainties. Comput
Math Appl 2013;65(1):89101.
[37] Helton JC, Johnson JD, Oberkampf WL. An exploration of alternative
approaches to the representation of uncertainty in model predictions. Reliab
Eng Syst Saf 2004;85(13):3971.
[38] Helton JC, Johnson JD, Oberkampf WL, Sallaberry CJ. Representation of analysis
results involving aleatory and epistemic uncertainty. Int J Gen Syst
2010;39(6):60546.
[39] Helton JC, Johnson JD. Quantication of margins and uncertainties: alternative
representations of epistemic uncertainty. Reliab Eng Syst Saf
2011;96(9):103452.
215
[40] Saltelli A, Chan K, Scott EM. Sensitivity analysis. New York, NY: Wiley; 2000.
[41] Helton JC, Johnson JD, Sallaberry CJ, Storlie CB. Survey of sampling-based
methods for uncertainty and sensitivity analysis. Reliab Eng Syst Saf
2006;91(1011):1175209.
[42] Saltelli A, Ratto M, Tarantola S, Campolongo F. Sensitivity analysis for chemical
models. Chem Rev 2005;105(7):281128.
[43] Martens H, Dijksterhuis GB, Byrne DV. Power of experimental designs,
estimated by Monte Carlo simulation. J Chemomet 2000;14(56):44162.
[44] Chen S, Cowan CFN, Grant PM. Orthogonal least squares learning algorithm for
radial basis function networks. IEEE Trans Neural Networks 1991;2(2):3029.
[45] Wasserman PD. Advanced methods in neural computing. New York: Van
Nostrand Reinhold; 1993.
[46] Firat M, Turan ME, Yurdusev MA. Comparative analysis of neural network
techniques for predicting water consumption time series. J Hydrol
2010;384(12):4651.
[47] Obrzud RF, Vulliet L, Truty A. Optimization framework for calibration of
constitutive models enhanced by neural networks. Int J Numer Anal Methods
Geomech 2009;33(1):7194.